Edge-to-Edge Achieved Transfer Throughput Inference Using Link Utilization Counts

We propose a methodology to infer edge-to-edge achieved transfer throughput using link utilization counts. Our method treats variations in the link utilization time-series as possible transfer starting or ending events. Iteratively following these variations to the neighboring routers, we then identify the path the transfer traversed through the monitored network. Our evaluation shows that this method can identify events larger than 3 Mbit/sec and longer than 2 minutes in duration with more than 95% recall. Additionally, we show that event detection is strongly correlated with the traffic in the busiest router in the path. We discuss how a number of applications such as throughput prediction and DDoS attack source detection can use the inferred information. Keywords-throughput inference; edge-to-edge; SNMP; network performance monitoring.